More Evidence That We Need More Evidence for AI Interventions

William Hersh, MD, Professor and Chair, OHSU
Blog: Informatics Professor
LinkedIn: William Hersh, MD
X: @williamhersh

In a previous post, I related the case of an excellent model that predicted hospital readmission yet when used in the context of real-world effort to reduce admissions was not able to lower the rate.

Some new studies highlight this scenario again of excellent models and systems that, when studied, do not show real-world benefit. A couple papers in Annals of Internal Medicine find a similar scenario for one of the earliest uses of artificial intelligence (AI) to demonstrate success, which is computer-aided detection (CADe) of polyps during colonoscopy results. A systematic review of previous clinical trials found that while there was an increased in detection of pre-cancerous adenomas but not of advanced adenomas and in higher rates of unnecessary removal of non-neoplastic polyps.[1]

The journal also featured a new randomized controlled trial (RCT) that showed no significant difference in advanced colorectal neoplasia detection rate (34.8% with intervention vs. 34.6% for controls) or mean number of advanced colorectal neoplasias detected per colonoscopy.[2]

An accompanying editorial notes the challenges in implementing AI in real world, which may impact RCT results, but we must build evidence base to support use.[3]

On a different clinical topic of predicting future trajectories in estimated glomerular filtration rate (eGFR) in adults with type 2 diabetes and chronic kidney disease, a new study in JAMA Network Open found that the new model excels over previous models in more accurate estimation of risk earlier in the disease course.[4] However, an accompanying editorial notes that while this model provides more accuracy, the benefit to those in this phase of the disease might be outweighed by “inappropriate avoidance of intravenous contrast, patient anxiety, and unnecessary testing with its associated costs.”[5] What is really needed, the author notes, are clinical trials to validate use of the model.

The research into these clinical applications of AI is important, and we must carry out the “basic science” research of them. But then we must move on to the next step of clinical application and studies that evaluate such systems in clinical trials or other appropriate evaluation methods.

For cited references in this article, see original source. Dr. Hersh is a frequent contributing expert to HealthIT Answers.